In the electronics industry, introducing Machine Learning (ML)-based
techniques can enhance Technology Computer-Aided Design (TCAD) methods.
However, the performance of ML models is highly dependent on their training
datasets. Particularly in the semiconductor industry, given the fact that the
fabrication process of semiconductor devices is complicated and expensive, it
is of great difficulty to obtain datasets with sufficient size and good
quality. In this paper, we propose a strategy for improving ML-based device
modeling by data self-augmentation using variational autoencoder-based
techniques, where initially only a few experimental data points are required
and TCAD tools are not essential. Taking a deep neural network-based prediction
task of the Ohmic resistance value in Gallium Nitride devices as an example, we
apply our proposed strategy to augment data points and achieve a reduction in
the mean absolute error of predicting the experimental results by up to 70%.
The proposed method could be easily modified for different tasks, rendering it
of high interest to the semiconductor industry in general.